Samplinglib
Lean gate passed 2026-08-19T05:09:39.794721+00:00 · 644be936998e
Chapter 11 · Book pp. 272–282 · August 9, 2026 edition

Non-Log-Concave Sampling

Use Fisher information to formulate and analyze approximate stationarity for non-log-concave targets, then compare upper and lower bounds.

Begin with 11.1 Open this chapter in the canonical August 9 source ↗

Chapter route

This chapter develops Fisher information, first-order stationarity, nonconvex Langevin dynamics, dissipativity. Its main destination is to connect the definitions below to the results that later chapters consume.

Core definitions

  • Relative Fisher information measures the squared score mismatch against the target.
  • Approximate first-order stationarity is formulated through small Fisher information rather than small sampling distance.
  • Dissipativity controls the drift outside a bounded region without global convexity.

Main results

  • Small Fisher information provides a meaningful stationarity certificate for non-log-concave targets.
  • Langevin and related algorithms admit finite-time Fisher-information bounds under stated regularity and moment assumptions.
  • Fisher-information control yields consequences for optimization and sampling observables.
  • Lower bounds delimit the attainable stationarity rates.

Contents

  1. 11.1Approximate First-Order Stationarity via Fisher InformationBook p. 272
  2. 11.2Fisher Information BoundsBook p. 274
  3. 11.3Applications of Fisher Information BoundsBook p. 276
  4. 11.4Lower BoundsBook p. 279
  5. 11.bibBibliographical NotesBook p. 280
  6. 11.exExercisesBook p. 280
Why is this chapter route valid?

Analytic contracts

  • Relative Fisher information needs an absolutely continuous law and a chosen score representative.
  • Entropy dissipation must be justified for the actual process and function domain, not only calculated formally.
  • Randomized-time output requires measurability of the time-indexed law and Fisher-information integrand.
  • Finite-time discretization and score-error bounds must specify the law under which every squared error is integrated.
  • Small Fisher information is a stationarity certificate, not automatically small total variation or rapid multimodal mixing.

Open boundaries

  • Relative-score representative and Fisher-information API
  • Rigorous entropy-dissipation/domain packet
  • Langevin discretization-to-Fisher-information chain
  • Book-specific applications and lower bounds
View Lean formalization

These mappings are evidence links, not a claim that the entire chapter is formalized.